MCP 服务器

ExcaliDraw Skill Pack

The diagram-quality layer for AI agents — an opinionated Excalidraw methodology (isomorphism test, evidence artifacts, multi-zoom, container discipline) plus a render-view-fix loop. Accepts skeleton and Mermaid input, 5 publishable themes, dual Node/Python renderer. Proven on a 77-diagram published book.

README

<h1 align="center">excalidraw-skill-pack</h1>

Diagrams that argue — not boxes that label. The diagram-quality layer for AI agents: an opinionated Excalidraw methodology + themes that turns "draw me a diagram" into figures that actually teach. Works in Claude Code, Cursor, GitHub Copilot, Codex, Gemini CLI, and any MCP-compatible agent.

Install

For	Command
Claude Code	`npx @excalidraw-skill-pack/install claude-code`
Cursor	`npx @excalidraw-skill-pack/install cursor`
GitHub Copilot	`npx @excalidraw-skill-pack/install copilot`
Codex	`npx @excalidraw-skill-pack/install codex`
Gemini CLI	`npx @excalidraw-skill-pack/install gemini-cli`
Any MCP agent	`npx @excalidraw-skill-pack/mcp-server`
Renderer only (Node)	`npx @excalidraw-skill-pack/render diagram.excalidraw --theme stripe-press`
Renderer only (Python)	`pipx install excalidraw-skill-pack-render && excalidraw-render diagram.excalidraw --theme stripe-press`

Why this, not just another Excalidraw MCP?

Excalidraw's official MCP and most community servers solve plumbing: get an agent to emit valid JSON and render it. They produce correct boxes-and-arrows.

This pack solves taste: an opinionated methodology — the isomorphism test, evidence artifacts, multi-zoom architecture, container discipline, one-accent-per-diagram — that decides what to draw and why, so the output is a visual argument that teaches instead of a labeled grid. That's the part a weekend of MCP code can't copy, and it's why the book gallery looks the way it does.

Plumbing is a commodity now. Quality isn't.

Themes

Five themes ship in v0.1. Authoring a new theme is 20 lines of JSON + npm publish:

npx @excalidraw-skill-pack/create-theme my-brand
cd theme-my-brand && npm publish --access public

Browse the theme registry →

Published packages (v0.1)

Channel	Package	Purpose
npm	`@excalidraw-skill-pack/core`	Methodology + bundled `default-sketchy` theme + schemas
npm	`@excalidraw-skill-pack/render`	Node renderer (`excalidraw-render` CLI)
npm	`@excalidraw-skill-pack/mcp-server`	stdio MCP server with 5 tools
npm	`@excalidraw-skill-pack/install`	One-command adapter installer
npm	`@excalidraw-skill-pack/create-theme`	Scaffolder for new theme packages
npm	`@excalidraw-skill-pack/theme-stripe-press`	Editorial / book-grade
npm	`@excalidraw-skill-pack/theme-notion`	Rounded, off-white
npm	`@excalidraw-skill-pack/theme-whiteboard`	Low-fi, bright, sketchy
npm	`@excalidraw-skill-pack/theme-dark`	Inverted contrast
PyPI	`excalidraw-skill-pack-render`	Python renderer (`excalidraw-render` CLI)

Proof

This skill drew 77 diagrams for the published technical book From Copilot to Colleague — argument spines, chapter openers, concept figures, and inline explainers, all from the methodology in this repo. No other Excalidraw generator can point at a corpus like that. A curated set — the book's argument spine plus all ten chapter openers — ships as source .excalidraw files in examples/book; the full 77 are live across fromcopilottocolleague.com.

Methodology

Diagrams are arguments. The shape should BE the meaning.

Isomorphism Test: would the structure alone communicate the concept?
Evidence artifacts: real code snippets, actual API names, concrete formats — not placeholder text.
One accent per diagram. Two means a competing argument; split it.

Read the full methodology (it's also what the AI agent reads).